DLMI_segmentation

This is the repository of the "Efficient 3D liver segmementation" project for the "CentraleSupelec - Spring 2020 MVA-DLMI: Deep Learning in Medical Imaging" course. In this project we tested a 3D version of Deeplab with a mobile net and a resnet backbone and compare it to a U-net and a attention gated Unet

The group of students was composed of three students :

• Emmanuel Jehanno
• Simon Brandeis
• Brice Rauby

The report is available on demand.

Prerequisites

Python3

Installation

1. Download the data from the LiTS challenge (train batch 1 and train batch 2)
2. Clone this repository
3. Run pip install -r requirements.txt

Getting Started

All the commands are to be executed from the main directory of this repository.

Training

• From the configs directory, choose the json file you want to train or create a new one and adapt the datapath to the folder containing the downloaded directories Training Batch 1 and Training Batch 2 .
• To launch a training run python utils/train.py --config_file=<path to config.json> --logdir=<path to the directory containing all the log dirs>

Evaluation

To evaluate a run use : python utils/error_analysis.py –run_dir=<path to the log dir> (the log dir is the one named after the date and the time of the training)

To run a the test time evalutation, run: python utils/speed_test.py